PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Zijian Chen, Xueguang Ma, Shengyao Zhuang, Jimmy Lin, Akari Asai, Victor Zhong

arXiv:2603.04384v3cs.CL

TL;DR

Deep Research agents expose reasoning traces that retrievers do not exploit, while their multi-turn subqueries lack tailored training data. AgentIR jointly embeds reasoning with queries and synthesizes training pairs; on BrowseComp-Plus, it reaches 68% end-to-end accuracy with Tongyi-DR versus 52% for a conventional embedding model twice its size.

  • Problem

    Deep Research agents reveal search intent in reasoning traces that existing retrievers do not exploit, and tailored training data for their multi-turn subqueries is lacking.

    Source: §1 · p. 2, §3.3 · p. 5

  • Method

    AgentIR jointly embeds an agent’s reasoning trace with its query and uses DR-Synth to create agent sub-query relevance pairs from standard QA datasets.

    Source: §1 · p. 2

  • Results

    AgentIR-4B shows consistent gains across multiple LLM agents; with Tongyi-DR, it achieves 68% end-to-end accuracy on BrowseComp-Plus versus 52% for a strong conventional embedding model twice its size.

    Source: §1 · p. 2, §6 · p. 12

  • Takeaways & Limitations

    Reasoning traces can ground searches in historical context and implicitly filter incorrect prior hypotheses, whereas explicitly embedding more uncurated history underperforms.

    Source: §6 · p. 12

Abstract

from arXiv · show

Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68\% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50\% with conventional embedding models twice its size, and 37\% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

1 Introduction

AgentIR jointly embeds Deep Research agents’ reasoning traces and queries, and DR-Synth creates tailored retrieval training pairs from standard QA data. Together they yield AgentIR-4B, which reaches 68% accuracy on BrowseComp-Plus with Tongyi-DeepResearch.

1 Introduction

  • Agent reasoning reveals search intent and evolving context, including prior findings, unresolved gaps, and hypotheses about promising search targets.In the example, prior searches identify the Grammy, while the trace hypothesizes progressive house as the subgenre with a “euphoric finale.”Source: §1 · p. 2
  • Reasoning-Aware Retrieval jointly embeds agent reasoning traces and queries, while DR-Synth creates agent sub-query–relevance training pairs from standard QA datasets.These contributions address retrievers’ failure to use reasoning context and the lack of training data for agent-issued sub-queries.Source: §1 · p. 2
  • 68% end-to-end accuracy on BrowseComp-Plus with Tongyi-DeepResearch, versus 52% for a conventional embedding model twice its size and 37% for BM25.AgentIR-4B also outperforms LLM-based reranking by 10% absolute.Source: §1 · p. 2
  • AgentIR-4B reduces the number of search steps needed and adds no inference overhead over query-rewriting methods because the reasoning traces are already generated.Source: §1 · p. 2
  • Without additional training, AgentIR-4B’s gains generalize to agents with different reasoning patterns, including gpt-oss-120B and GLM-4.7.Source: §1 · p. 2
  • Reasoning traces improve retrieval by summarizing relevant earlier findings and implicitly filtering outdated or incorrect information.Source: §1 · p. 2

2 Related Work

Related work situates Reasoning-Aware Retrieval in the shift to multi-turn Deep Research and distinguishes it from both reasoning-focused retrieval and methods for resolving ambiguous human queries. Its key signal is the agent’s reasoning trace, which articulates intent during collaborative, multi-turn search.

Deep Research Agents

Retrieval and Reasoning

  • Unlike ReasonIR and RaDeR, which target reasoning-intensive tasks without an agent, Reasoning-Aware Retrieval supports collaborative multi-turn retrieval with an agent (Shao et al., 2025; Das et al., 2025).The agent interleaves explicit reasoning and retrieval while resolving complex tasks.Source: §2 · p. 3

Understanding Ambiguous Queries

  • Prior ambiguity methods add human instructions (Asai et al., 2023), ask clarifying questions (Aliannejadi et al., 2019), or use LLM-generated hypothetical context (Gao et al., 2023); here, the agent’s reasoning trace articulates intent without human annotation.Ambiguous-query handling has long been studied in information retrieval.Source: §2 · p. 3

3 Methodology

The methodology models Deep Research as a multi-turn agent whose reasoning traces and searches reflect interaction history. It uses those traces to augment retrieval and synthesizes sub-query training data from standard QA datasets.

3.1 Preliminary

  • Each turn, a ReAct-style agent conditions a reasoning trace and action on interaction history, then searches or ends with a final answer.The action is a search call or a final answer that terminates the loop.Source: §3.1 · p. 3

3.2 Reasoning-Aware Retrieval

  • Reasoning-Aware Retrieval jointly embeds the agent’s reasoning trace and query, using context that clarifies intent, reflects prior results, and grounds hypotheses in interaction history.Unlike task-aware retrieval requiring human instructions or HyDE’s additional query-expansion LLM call, the agent produces its trace as part of its normal loop.Source: §3.2 · p. 4

3.3 DR-Synth: Constructing Training Data for Deep Research Queries

  • Standard QA datasets lack relevance supervision for local multi-turn sub-queries, so DR-Synth generates sub-query-level retriever training data from QA datasets using agent rollouts.Existing positive documents support the global question, while a local sub-query may target only part of it.Source: §3.3 · p. 5
  • For each QA triple, DR-Synth runs an agent with a query-only retriever, extracts reasoning-query pairs from search turns, and trains only on rollouts that answer the question successfully.A trajectory with T−1 search turns yields T−1 training instances per original question.Source: §3.3 · pp. 5–6
  • DR-Synth labels each turn by reranking query-only retrieval candidates against the local query and global question-answer pair, using the top-ranked document as positive and the bottom seven as hard negatives.The candidate pool includes the top 50 retrieved documents and the QA triple’s positive documents; an LLM performs listwise reranking.Source: Figure 2 · p. 4, §3.3 · pp. 5–6

4 Experiments

On BrowseComp-Plus, AgentIR-4B achieves the best performance across evaluated agents, with gains in accuracy and search efficiency. The experiments also assess its training setup, evaluation protocol, baselines, and transfer across agents and tools.

4.1 Training Setup

  • AgentIR-4B is trained by applying DR-Synth to WebShaper triples, yielding 5,238 instances used to contrastively fine-tune Qwen3-Embedding-4B.Rollouts use Tongyi-DeepResearch as the agent and Qwen3-Embedding-8B as the query-only retriever.Source: §4.1 · p. 6

4.2 Evaluation Setup and Metrics

  • Evaluation pairs three open-weight Deep Research agents with retrievers on BrowseComp-Plus, whose multi-hop queries require 20+ searches; agents receive top-5 snippets truncated to 512 tokens.Tongyi-DR is also evaluated with and without a visit tool that opens full documents.Source: §4.2 · p. 6
  • The evaluation reports end-to-end QA Accuracy, recall of retrieved ground-truth evidence documents, and the number of Search Calls before the agent answers.Source: §4.2 · p. 6

4.3 Baselines

  • Baselines include query-only Qwen3 embeddings, BM25 and ReasonIR-8B; query expansion with Reason-Rewriter + Reason-Embed-8B and Agentic-R; and Qwen3-8B reranking of Qwen3-Embedding-4B’s top 20 results.The query expansion and reranking methods provide comparisons to alternative ways of refining retrieval.Source: §4.3 · pp. 6–7

4.4 Results

  • 66.27% accuracy: AgentIR-4B leads all evaluated retrievers across agents on BrowseComp-Plus, improving 17.60% absolute over Qwen3-Embed-4B and ≈15% absolute over twice-size Qwen3-Embed-8B.The reported 17.60% absolute improvement is over the Qwen3-Embed-4B backbone.Source: Table 1 · p. 7, §4.4 · pp. 7–8
  • AgentIR-4B reduces search calls from 32.92 with BM25 to 25.91 and achieves approximately 10% higher absolute accuracy than computationally expensive Qwen3-Embed-4B + LLM Rerank, despite performing no reranking.Source: §4.4 · p. 8
  • AgentIR-4B surpasses past single-turn reasoning-intensive retrievers such as ReasonIR-8B by ≈15% accuracy; HyDE-like expanders appear ineffective in Deep Research settings, while AgentIR-4B’s access to agent reasoning yields more grounded hypotheses.Source: §4.4 · p. 8
  • AgentIR-4B generalizes zero-shot from WebShaper queries to BrowseComp-Plus, transfers from Tongyi-DR trajectories to gpt-oss-120B and GLM-4.7 without fine-tuning, and retains gains with Tongyi-DR’s visit tool.The passage attributes transfer to agents with distinct reasoning styles and search behaviours.Source: §4.4 · p. 8

5 Analysis

The analysis finds that reasoning-aware retrieval and DR-Synth each contribute to retrieval accuracy, while adding broader trajectory history offers little benefit and can introduce noise. The current reasoning provides a useful, comparatively curated retrieval signal.

5.1 Effectiveness of Reasoning-Aware Retrieval and DR-Synth

  • On Tongyi-DR, reasoning-only raises accuracy from 48.67% to 55.54%, DR-Synth training without reasoning reaches 59.4%, and combining both reaches 66.27%.Both components are independently effective, and their combination gives the largest gains; the pattern holds across all three agents.Source: §5.1 · p. 8, Table 2 · p. 9

5.2 Alternative Sources of Retrieval Signals

  • AgentIR-4B outperforms Prior Queries and Global Question, while adding reasoning to Prior Queries helps but still underperforms AgentIR-4B.The comparison shows reasoning traces provide a distinct retrieval signal beyond prior queries or the global question.Source: Table 3 · p. 10, §5.2 · p. 10
  • For Tongyi-DR, full-history retrieval has zero recall in 11.45% of runs, which average 37.46 search turns versus 27.18 overall.The authors attribute the modest or negative results to noisy retrieved documents propagating errors into later queries.Source: §5.2 · p. 10

5.3 Effect of Adding Prior Reasonings

  • On Tongyi-DR, adding prior turns beyond the current reasoning does not improve end-to-end accuracy.The comparison evaluates history lengths k = 1, 2, 5, 9, 17, and all; k = 1 corresponds to AgentIR-4B.Source: Figure 3 · p. 11, §5.3 · p. 11
  • The current reasoning alone covers more than 40% of all past clues, with diminishing coverage gains as more prior reasonings are included.This is consistent with current reasoning often summarizing findings from prior turns.Source: Figure 3 · p. 11, §5.3 · pp. 11–12
  • Earlier reasoning can add noise: as more prior turns are included, incorrect claims increase substantially, while current reasoning can summarize confirmed findings and filter outdated guesses.The paper illustrates this with discarded candidate countries and artists, and reports measuring correct and incorrect claims across recent reasonings.Source: Figure 4 · p. 11, §5.3 · p. 12

6 Conclusion · A Main AgentIR-4B Prompt

AgentIR combines reasoning-aware retrieval, which uses agents’ explicit reasoning traces, with DR-Synth, which creates multi-turn retriever training data from standard QA datasets. The paper reports that both components are independently effective and that their combination yields consistent gains across multiple LLM agents.

6 Conclusion

  • AgentIR-4B combines Reasoning-Aware Retrieval and DR-Synth, with consistent gains across multiple LLM agents over existing retrievers.The paper reports that each component is independently effective.Source: §6 · p. 12
  • Reasoning-Aware Retrieval uses agents’ explicit natural-language reasoning traces as contextual intent signals absent from traditional human search.Unlike conventional retrievers that process isolated queries, it incorporates the agent’s reasoning alongside its query.Source: §6 · p. 12
  • DR-Synth addresses the lack of multi-turn retriever training data for Deep Research by synthesizing data from standard QA datasets.Source: §6 · p. 12
  • Reasoning traces ground each search in historical context and implicitly filter incorrect prior hypotheses; embedding more uncurated history underperforms.Source: §6 · p. 12
  • The findings motivate retrieval-focused context engineering: principled curation methods that optimize retrievers’ view of an evolving problem.Source: §6 · p. 12
  • Agent-optimized contexts may be directly reused by retrievers as agent context engineering advances, improving performance without additional computational cost.Source: §6 · p. 12
  • The authors anticipate humans increasingly delegating complex information seeking to agents and release AgentIR-4B to encourage research for agent users.Source: §6 · p. 12

A Main AgentIR-4B Prompt

  • Figure 5’s AgentIR-4B prompt fills {reasoning} with τt and {query} with qt at turn t; its duplicate “Query:” is intentional for Qwen3-Embedding-4B’s instruction format (Zhang et al., 2025).Source: Figure 5 · p. 17

B AgentIR-4B Training Details

AgentIR-4B is trained by contrastively fine-tuning a reasoning-and-query embedding model on DR-Synth data generated from WebShaper. The training pipeline builds a retrieval corpus and produces 5,238 instances from correctly answered rollouts.

B AgentIR-4B Training Details

  • AgentIR-4B fine-tunes Qwen3-Embedding-4B (Zhang et al., 2025) with LoRA (Hu et al., 2022) on contrastive loss, embedding concatenated reasoning traces and queries with DR-Synth positives and in-batch negatives.Source: B AgentIR-4B Training Details · p. 17
  • Training uses a 1e-4 learning rate, batch size 4, maximum document length 4096, maximum query length 8192, and two epochs on DR-Synth-generated WebShaper data, using Tevatron (Ma et al., 2025) on one H100.Gradient checkpointing and gradient accumulation of 2 are used.Source: B AgentIR-4B Training Details · p. 17
  • DR-Synth rollouts use 500 WebShaper (Q, A, P) pairs, with Tongyi-DR as agent, Qwen3-Embedding-8B as retriever, and GLM-4.6 as oracle reranker.At each retrieval turn, the retriever returns 50 documents for reranking; the top-ranked document is labeled positive, and the top 5 are returned to continue the rollout.Source: B AgentIR-4B Training Details · p. 17
  • Because WebShaper supplies only positive-document URLs, positives are scraped with Selenium and parsed with Trafilatura to build the training corpus.Source: B AgentIR-4B Training Details · p. 17
  • Hard negatives come from roughly 6 GPT-4o-generated sub-queries per question, with up to 100 SerpAPI results scraped per sub-query; one million FineWeb documents (Penedo et al., 2024) are added as random negatives.Source: B AgentIR-4B Training Details · p. 17
  • 250 of 500 WebShaper queries were answered correctly, yielding 5,238 training instances from correct trajectories and a deduplicated corpus of 1,146,942 documents.Source: B AgentIR-4B Training Details · p. 17

C Prompts for Alternative Sources of Retrieval Signals

Alternative retrieval signals use the same training procedure, changing the retriever input to variant-specific transformations of agent history formatted with prompts.

D Prompt for Adding Prior Reasonings · E Prompts for Atomic Clues

Appendix D describes training and prompt formatting for adding prior reasoning turns, while Appendix E identifies prompts for decomposing reasonings into atomic clues and assigning those clues to documents.

D Prompt for Adding Prior Reasonings

E Prompts for Atomic Clues

  • Figure 10 shows the prompt for decomposing reasonings into atomic clues.The passage groups this prompt with the document-assignment prompt in Figure 11.Source: E Prompts for Atomic Clues · p. 19
  • Figure 11 shows the prompt for assigning atomic clues to documents.The passage pairs it with the reasoning-decomposition prompt in Figure 10.Source: E Prompts for Atomic Clues · p. 19

F HyDE Example

This section presents a full example of Reason-Rewriter’s HyDE-style query expansion.

  • The full example of Reason-Rewriter’s HyDE-style expansion for the query “backroom studio early 2010s euphoric” is shown in Figure 12.Source: F HyDE Example · p. 19

G Prompt for Noise Analysis

To identify noisy claims, the authors replace direct LLM judgment with a two-step process: extract intermediate answers from evidence documents, then classify claims using a ground-truth answer list that includes those answers.

G Prompt for Noise Analysis

  • The two-step process first extracts intermediate answers for each hop from the question, answer, and full evidence documents, then labels claims against a ground-truth answer list.The list includes the intermediate answers; the extraction and classification prompts are shown in Figures 13 and 14, respectively.Source: G Prompt for Noise Analysis · p. 19, Figure 13 · p. 22, Figure 14 · p. 22
  • Directly asking an LLM to judge claims using only the question-answer pair is inaccurate because it lacks intermediate hops and can mislabel correct intermediate hypotheses.Source: G Prompt for Noise Analysis · p. 19
  • The extraction step uses the question, answer, and full evidence documents required to answer the question to recover each hop’s intermediate answer.Source: G Prompt for Noise Analysis · p. 19, Figure 13 · p. 22

Related papers

Loading 2603.04384v3…